Multifidelity Simulation-Based Inference

Updated 9 December 2025

Multifidelity simulation-based inference schemes are advanced methods that combine a hierarchy of low- and high-fidelity simulators to achieve unbiased, cost-efficient statistical inference.
They integrate Bayesian frameworks, Monte Carlo techniques, surrogate modeling, and neural density estimation to optimize resource allocation and reduce computational costs.
Empirical benchmarks in fields like engineering, physics, and biology report 3×–100× reductions in simulation costs without sacrificing accuracy.

Multifidelity simulation-based inference schemes form a diverse and rigorously analyzed class of methodologies that exploit hierarchies of simulators—ranging from inexpensive, low-fidelity surrogates to costly, high-fidelity physical or stochastic models—to produce accurate statistical inference at a fraction of traditional computational cost. These methods are mathematically grounded, span both Bayesian and likelihood-free frameworks, and incorporate techniques from Monte Carlo, surrogate modeling, Gaussian processes, and modern neural density estimation.

1. Mathematical Foundations and Core Principles

The central paradigm is to substitute, combine, or hierarchically fuse outputs from models at varying fidelities in order to maintain or improve inference accuracy relative to single-fidelity approaches, while significantly reducing evaluation cost. This is achieved with various theoretical constructs:

Telescoping Series and Randomization: Certain methods, such as the multi-fidelity pseudo-marginal MCMC, represent the exact target density π₍∞₎(θ) as a telescoping sum of incrementally more accurate low-fidelity models, $\pi_\ell(\theta)$ , with increments $\Delta_\ell(\theta)$ (Cai et al., 2022). An unbiased estimator is obtained by random truncation (“Russian roulette”) with weights $w_\ell$ , such that

$\hat\pi(\theta) = \sum_{\ell=0}^L \frac{\Delta_\ell(\theta)}{w_\ell}$

The estimator preserves unbiasedness: $\mathbb{E}[\hat\pi(\theta)] = \pi_\infty(\theta)$ .

Surrogate Model Correction and Fusion: Approaches for rare event simulation (e.g., engineering reliability analysis) construct corrected low-fidelity surrogates using Gaussian Process (GP) models of the discrepancy, and assemble a multifidelity surrogate via model averaging, probabilistic selection, or active learning (Chakroborty et al., 2022).
Multifidelity MLMC Estimators: Extensions of classical Multilevel Monte Carlo (MLMC) replace or supplement high-fidelity samples with control variate differences between adjacent fidelity levels, variance-optimal allocation of resources, and coupling strategies to maximize statistical correlation (Muchandimath et al., 20 Oct 2025, Hikida et al., 6 Jun 2025).
Likelihood-Free Frameworks and ABC: Multifidelity Approximate Bayesian Computation (ABC) leverages early accept/reject schemes, where cheap low-fidelity outputs are used to screen, shortcut, or reweight rare high-fidelity simulations, with mathematically derived unbiased weights and explicit cost/variance trade-offs (Prescott et al., 2018).
Bayesian Model Fusion and Neural Inference: Neural simulation-based inference methods generalize transfer learning, knowledge distillation, and MLMC objectives to leverage multifidelity simulation datasets within a unified probabilistic framework. These include neural posterior estimation (NPE) with transfer/fine-tuning, feature-matching between fidelity levels, and MLMC-structured loss aggregation (Krouglova et al., 12 Feb 2025, Thiele et al., 1 Jul 2025, Hikida et al., 6 Jun 2025, Saoulis et al., 27 May 2025).

2. Methodological Taxonomy

Different classes of multifidelity schemes have emerged for simulation-based inference problems:

(A) Pseudo-Marginal MCMC with Randomized Fidelity

Employs a randomized, unbiased estimator of the target posterior by random truncation in a telescoping expansion across fidelity levels.
Embeds the estimator in a pseudo-marginal MCMC with extended state including auxiliary randomness, preserving the exact target as invariant.
Theoretical guarantees: unbiasedness, ergodicity, variance/cost optimality with tunable truncation weights (Cai et al., 2022).

(B) Multifidelity Surrogate Models in Rare Event Analysis

Constructs corrected surrogate models for each low-fidelity code via GPs, then fuses predictions using locally determined model probabilities that account for both predictive accuracy and cost.
Active learning strategies within subset simulation framework decide on-the-fly which models to use and when to request expensive high-fidelity queries, yielding drastic reduction in required HF evaluations (Chakroborty et al., 2022, Dhulipala et al., 2021).

(C) Adaptive Resource Allocation in Likelihood-Free Bayesian Inference

Multifidelity importance sampling and ABC approaches allocate simulation effort adaptively across fidelity levels using analytic or online estimates of cost, variance, and cross-model correlation.
Optimal allocation balances exploration (estimating cross-moments) and exploitation (sampling surrogates that yield greatest error reduction per cost), attaining provably optimal MSE/budget scaling (Prescott et al., 2021, Han et al., 2023).

(D) Multifidelity Neural Simulation-Based Inference

Pretraining neural surrogates on large low-fidelity datasets, followed by fine-tuning on sparse high-fidelity data (transfer learning).
MLMC loss construction integrates telescoping differences of neural network outputs across fidelities, yielding lower-variance gradient estimates and improved accuracy under fixed simulation budgets (Krouglova et al., 12 Feb 2025, Hikida et al., 6 Jun 2025, Thiele et al., 1 Jul 2025, Saoulis et al., 27 May 2025).
Advanced schemes apply feature-matching and response distillation to align posteriors across fidelities in high dimension (Thiele et al., 1 Jul 2025).

3. Theoretical Properties and Guarantees

Unbiasedness and Exactness: Pseudo-marginal MCMC and MLMC telescoping-sum estimators are provably unbiased for the high-fidelity target; no bias is introduced by randomization, estimator design, or control variate fusion (Cai et al., 2022, Muchandimath et al., 20 Oct 2025, Hikida et al., 6 Jun 2025, Prescott et al., 2021).
Variance-Optimality: Allocation of simulation effort (weights, number of calls per fidelity, adaptive early-accept/reject probabilities) is derived to minimize the asymptotic variance at a fixed computational budget. For ABC, explicit closed-form expressions for optimal continuation probabilities are proven (Prescott et al., 2018, Prescott et al., 2021).
Resource Efficiency: Across both theoretical analysis and empirical validation, multifidelity schemes demonstrate $3\times$ – $100\times$ reduction in wall-clock time or simulation cost at fixed accuracy, depending on the simulation task and cross-fidelity correlation (Cai et al., 2022, Krouglova et al., 12 Feb 2025, Muchandimath et al., 20 Oct 2025, Chakroborty et al., 2022).

4. Algorithmic Structures and Pseudocode Insights

Typical multifidelity simulation-based inference algorithms combine:

Randomized Estimation: Russian roulette truncation or Poisson randomization for unbiasedness (Cai et al., 2022).
Active Learning/Acquisition: Sequential policies that select inputs and fidelity levels for simulation based on maximum expected uncertainty reduction per cost (MSUR), ensemble variance, or other acquisition functions (Stroh et al., 2017, Chakroborty et al., 2022, Krouglova et al., 12 Feb 2025).
Model Correction and Surrogate Fusion: GP discrepancy modeling and probabilistic model selection/averaging at prediction time (Chakroborty et al., 2022, Dhulipala et al., 2021).
Sequential and Adaptive Updating: Importance weights, continuation probabilities, or resource allocation parameters updated based on online estimates of underlying cross-model statistics (Prescott et al., 2018, Prescott et al., 2021, Han et al., 2023).
Gradient-Based Inference: For differentiable surrogates, backpropagation is facilitated entirely via low-fidelity adjoint codes, whereas high-fidelity solvers need only be black-box (Nitzler et al., 30 May 2025).

Table: Key Attributes of Representative Multifidelity Inference Schemes

Method	Model Fusion Paradigm	Theoretical Guarantee
MF pseudo-marginal MCMC (Cai et al., 2022)	Randomized telescoping sum	Unbiasedness, ergodicity
MF-ABC (early accept/reject) (Prescott et al., 2018)	Stochastic screening, unbiased weight	Optimal efficiency, no ABC bias
MLMC-SBI (Hikida et al., 6 Jun 2025, Muchandimath et al., 20 Oct 2025)	Control variate (difference coupling)	Variance/cost optimality, no bias
Feature-matching SBI (Thiele et al., 1 Jul 2025)	Probabilistic mapping + distillation	Consistent posterior, reduced budget
BMFIA (Nitzler et al., 30 May 2025)	Learned conditional density (P-CAE)	Full differentiability, high-dim HMC
MF Bayesian surrogate (Chakroborty et al., 2022)	GP correction, local model fusion	Provable COV reduction, adaptivity
Adaptive resource allocation (Prescott et al., 2021)	Piecewise-constant mean allocation	MSE optimality, adaptive allocation

5. Practical Applications and Empirical Benchmarks

Multifidelity simulation-based inference schemes demonstrate broad applicability:

Stochastic Kinetics: Bayesian inference for chemical reaction networks (CME) with multifidelity CMEs and adaptive fidelity selection yielding $3\times$ – $6\times$ time reduction (Catanach et al., 2020).
Rare Event Simulation in Engineering: Surrogate GP fusion and active learning enable estimation of small failure probabilities ( $P_f\sim10^{-4}$ – $10^{-6}$ ) with $>98\%$ reduction in high-fidelity simulation cost (Chakroborty et al., 2022, Dhulipala et al., 2021).
Computational Physics and CFD: In turbulence model calibration, transport map-based coupling between MCMC chains at different fidelities achieves $\sim50\%$ wall-clock savings without posterior degradation (Muchandimath et al., 20 Oct 2025).
Systems Biology: Multifidelity ABC for non-Markovian gene networks achieves $10$– $100\times$ speed-up versus standard ABC (Steele et al., 2 Dec 2025).
Cosmology and Scientific ML: Multilevel neural SBI, transfer learning, and knowledge distillation/deep mapping approaches reduce high-fidelity simulation budgets by one to two orders of magnitude while matching gold-standard inference accuracy (Krouglova et al., 12 Feb 2025, Saoulis et al., 27 May 2025, Thiele et al., 1 Jul 2025, Hikida et al., 6 Jun 2025).
High-Dimensional Inverse Problems: BMFIA in high-dimensional spatial field inference (e.g., poro-elasticity, Darcy flow) attains accurate posterior reconstruction using only LF adjoints and a handful of HF runs ( $\sim50$ –$300$), with $65\times$ – $158\times$ net speed-ups (Nitzler et al., 30 May 2025).

6. Limitations, Extensions, and Contemporary Directions

Assumptions on LF–HF Correlation: All multifidelity methods rely on the existence of sufficiently accurate, inexpensive surrogates. Poor low-fidelity surrogates limit the overall benefit; cost–accuracy trade-offs must be empirically quantified for each problem (Cai et al., 2022, Chakroborty et al., 2022).
Hyperparameter Sensitivity: Some approaches require tuning of truncation weights, model probabilities, or continuation rates; improper choice can reduce gains (Prescott et al., 2018, Krouglova et al., 12 Feb 2025).
Sequential/Adaptive Resource Allocation: Ongoing research investigates fully adaptive schemes, further automation of hyperparameter tuning, and integration of Bayesian model selection between fidelity levels (Prescott et al., 2021, Chakroborty et al., 2022).
Scalability: While some methods support inference in $d\gg100$ -dim parameter spaces via full differentiability (e.g., BMFIA), GP-based methods can encounter computational bottlenecks in very high dimension unless approximations or deep surrogates are incorporated (Nitzler et al., 30 May 2025, Chakroborty et al., 2022).
Integration with Modern ML: Unifying feature-matched probabilistic mappings, neural surrogate architectures, and advanced gradient-based variational inference is an active area, with proposals to combine with diffusion models, deep surrogates, and hierarchical learning (Thiele et al., 1 Jul 2025, Hikida et al., 6 Jun 2025, Krouglova et al., 12 Feb 2025, Saoulis et al., 27 May 2025).

7. Outlook and Conclusions

Multifidelity simulation-based inference schemes represent a mature and rapidly evolving methodology for statistical inference in computationally intensive scientific domains. They provide principled, unbiased, and cost-effective algorithms by leveraging hierarchies of models, optimal allocation strategies, advanced surrogate corrections, and neural network architectures. Empirical results across physics, engineering, neuroscience, cosmology, and systems biology consistently show that multifidelity schemes can accelerate inference by factors of $3\times$ to $100\times$ —without loss of accuracy relative to single-fidelity, high-resolution inference. Future work aims to further automate cross-fidelity knowledge transfer, enable robust deployment in ultra-high-dimensional settings, and extend multifidelity techniques to new classes of scientific simulation and data analysis problems.